Box CEO Aaron Levie on AI's 'Era of Context'

Box Launches BoxWorks with New AI-Powered Features

On Thursday, Box initiated its developer conference, BoxWorks, with the announcement of a new suite of AI features. These features integrate agentic AI models directly into the core of the company’s product offerings.

Accelerated AI Development at Box

The conference featured a greater number of product announcements than usual, reflecting the rapidly increasing pace of AI development within the company. Box previously launched its AI studio last year, followed by data-extraction agents in February, and further advancements for search and in-depth research in May.

Introducing Box Automate: An Operating System for AI Agents

The company is now introducing Box Automate, a new system designed to function as an operating system for AI agents. This system divides workflows into distinct segments, allowing for AI augmentation where necessary.

Interview with CEO Aaron Levie

A conversation with CEO Aaron Levie explored the company’s AI strategy and the challenges of competing with companies developing foundation models. He expressed strong optimism regarding the potential of AI agents in the contemporary workplace, while also acknowledging the current limitations of existing models and strategies for managing them with established technologies.

This interview has been condensed for brevity and clarity.

The Vision Behind Integrating AI Agents

You’re unveiling numerous AI products today. Could you elaborate on the overarching vision? Why incorporate AI agents into a cloud content-management service?

Our primary focus at Box revolves around understanding how AI is reshaping work. Currently, the majority of this impact is concentrated on workflows involving unstructured data. Automation of structured data workflows, those feeding into databases, has already been achieved. Systems like CRM, ERP, and HR have benefited from years of automation.

The Untapped Potential of Unstructured Data

However, automation has historically been absent in areas dealing with unstructured data. Consider legal review, marketing asset management, or M&A deal reviews – all heavily reliant on unstructured data. Previously, these processes required manual review, updates, and decision-making. Computers lacked the capability to effectively interpret documents or marketing materials.

AI agents now enable us to access and leverage this previously untapped potential within unstructured data.

Addressing Risks and Ensuring Reliability

What are the potential risks associated with deploying agents in a business environment? Are customers concerned about using this technology with sensitive data?

Customers are prioritizing consistent execution. They require assurance that each workflow run will yield similar results, at the same stage, without unpredictable deviations. Agents must avoid compounding errors that could escalate with repeated submissions.

Deterministic Guardrails and Agentic Flexibility

Establishing clear demarcation points is crucial, defining where the agent’s influence begins and ends. Each workflow requires a balance between deterministic guardrails and agentic, non-deterministic processes. Box Automate allows users to control the scope of each agent’s tasks before handing off to the next, enabling scalable AI agent deployment across various organizational workflows.

Mitigating Context Window Limitations

What challenges does splitting up the workflow address?

Even advanced systems like Claude Code exhibit limitations. Models eventually exhaust their context window, hindering their ability to make informed decisions. There are inherent constraints in AI; a perpetually running agent with unlimited context cannot effectively handle all business tasks. Breaking down workflows and utilizing subagents is therefore essential.

We are currently in an era where context is paramount for AI. AI models and agents require context, and that context resides within your unstructured data. Our system is designed to provide agents with the necessary context to maximize their performance.

Smaller vs. Larger Models: A Strategic Approach

There’s an ongoing debate regarding the advantages of large, powerful models versus smaller, more reliable ones. Does this position you in favor of smaller models?

It’s important to note that our system doesn’t restrict task complexity or length. Our goal is to provide the appropriate safeguards, allowing you to determine the level of agentic behavior for each task.

We don’t advocate for a specific approach. We are focused on creating a future-proof architecture that will seamlessly integrate improvements in models and agentic capabilities.

Data Control and Security

Data control is another concern. Models are trained on vast datasets, raising fears of sensitive data being revealed or misused. How does this factor into your approach?

Many AI deployments falter due to a perceived simplicity. The assumption that granting an AI model access to all unstructured data will automatically yield answers is often flawed. This can lead to the disclosure of data individuals shouldn’t access.

A robust layer managing access controls, data security, permissions, data governance, and compliance is essential. Our decades of experience building a system that addresses this exact problem – ensuring appropriate data access – is a significant advantage. When an agent answers a question, you can be certain it won’t draw on data the user is not authorized to view. This is fundamentally built into our system.

Competition from Foundation Model Companies

Anthropic recently released a feature for directly uploading files to Claude.ai. While different from Box’s file management capabilities, how do you view potential competition from foundation model companies?

Enterprises deploying AI at scale require security, permissions, and control. They need a user-friendly interface, powerful APIs, and the flexibility to choose the AI model best suited for each use case. They don’t want to be locked into a single platform.

We’ve built a system that provides all these capabilities. We handle storage, security, permissions, vector embedding, and connect to every leading AI model available.